Skip to content
certsim
About

Comparisons

The distinctions exam questions are built on.

15 results · content in English

SageMaker inference: real-time vs serverless vs async vs batch

Requirement / use caseChoose
A news site's article-recommendation model receives steady traffic all day and must respond in milliseconds.Real-time inference
A nonprofit's donation-risk model is called only a few times a week, and brief cold-start delays are acceptable.Serverless Inference
A media company submits very large video files that each take many minutes to analyze; requests should be queued.Asynchronous Inference
A bank scores its entire customer table once a month and does not need any endpoint afterward.Batch Transform
  • Start with latency, request shape, traffic pattern and whether a persistent endpoint is needed.

Managed AI services by use case

Requirement / use caseChoose
Convert recorded customer voicemails into text.Amazon Transcribe
Read news articles aloud in a natural-sounding voice.Amazon Polly
Determine whether product reviews are positive, negative, neutral, or mixed.Amazon Comprehend
Extract key-value pairs and tables from scanned tax forms.Amazon Textract
Read license plate numbers from traffic camera images.Amazon Rekognition
Build a text and voice bot that collects order details from customers.Amazon Lex
  • Identify the input and output modality before choosing a service.

Regression vs classification vs clustering

Requirement / use caseChoose
Estimate the delivery time, in minutes, for each online order.Regression
Predict whether a store-card applicant will default, using past applicants labeled as defaulted or not.Binary classification
Group shoppers into segments by purchasing behavior when no segment labels exist.Clustering
Flag unusual temperature readings from refrigeration units when no failures have ever been labeled.Anomaly detection
Predict daily demand for each product over the next 12 weeks from past daily sales.Time-series forecasting
  • A numeric target is regression; named categories are classification; unknown groups suggest clustering.

SageMaker lifecycle components

Requirement / use caseChoose
Have a private workforce draw bounding boxes on product images to create training labelsSageMaker Ground Truth
Visually import, clean, and transform tabular data and generate a data quality report with little codeSageMaker Data Wrangler
Store engineered features once and serve the same values for training and low-latency inferenceSageMaker Feature Store
Orchestrate processing, training, evaluation, and deployment steps as a repeatable workflowSageMaker Pipelines
Catalog model versions and track their approval status before deploymentSageMaker Model Registry
Detect drift in production data and model quality against a baselineSageMaker Model Monitor
  • Label creation, feature storage, workflow orchestration, registry and drift monitoring solve different problems.

AgentCore components

Requirement / use caseChoose
Host the agent on serverless infrastructure where each user session runs in its own isolated microVM.AgentCore Runtime
Remember each customer's preferences and important facts across future conversations.AgentCore Memory
Convert existing REST APIs and AWS Lambda functions into MCP-compatible tools for the agent.AgentCore Gateway
Manage OAuth tokens so that the agent can access third-party services, such as Slack, on behalf of users who signed in through the corporate identity provider.AgentCore Identity
Trace each step of the agent's workflow and view latency, token usage, and error metrics in Amazon CloudWatch.AgentCore Observability
Deterministically allow or deny specific tool calls based on user identity and tool input parameters.Policy in AgentCore
  • Gateway exposes tools; Identity supplies credentials; Policy authorizes calls; Memory stores context.

Bedrock inference and cost controls

Requirement / use caseChoose
A prototype with low, unpredictable traffic and no appetite for commitments.On-demand inference
Summarizing a large archive of documents overnight; results are not needed until the next day.Batch inference
A steady, high-volume production application that needs a guaranteed throughput level at a fixed hourly price.Provisioned Throughput
A chatbot that sends the same long policy document at the start of every request.Prompt caching
  • Choose by workload; availability and pricing vary by model and Region.

RAG vs fine-tuning vs continued pre-training vs distillation

Requirement / use caseChoose
Learns broad language patterns from massive amounts of unlabeled data, starting without any task-specific knowledgePre-training
Further trains an existing FM on a large unlabeled corpus from one domain so that it absorbs that domain's vocabularyContinued pre-training
Trains an existing FM on labeled prompt-response examples to improve performance on a specific taskFine-tuning
Uses a larger model's responses to train a smaller, faster model for a specific use caseDistillation
  • RAG supplies knowledge at inference time; training techniques change weights.

Temperature vs top P vs top K

Requirement / use caseChoose
Make headlines more varied and creative by increasing the likelihood of lower-probability tokens.Temperature
Limit the model to choosing among the 40 most likely next tokens.Top K
Limit sampling to tokens within the top 90% of the cumulative probability distribution.Top P
End generation as soon as the model outputs the marker "</summary>".Stop sequences
Cap each generated summary at 200 tokens to control output cost.Maximum response length (max tokens)
  • Sampling controls do not guarantee truth or determinism; max tokens controls response length.

Evaluation metrics

Requirement / use caseChoose
Score machine translations of product descriptions against professional reference translations.BLEU
Measure how much of an editor's reference summary is recovered by a generated summary, based on shared n-grams.ROUGE
Score generated answers against reference answers so that paraphrases with the same meaning still score high.BERTScore
Score short extractive answers by the harmonic mean of word-level precision and recall against the ground truth.F1 score
Rate open-ended chatbot replies for helpfulness when no reference answers exist, with an explanation for each rating.LLM-as-a-judge
  • Choose metrics that fit the task; reference similarity alone does not prove factuality or business value.

RAG retrieval and answer evaluation

Requirement / use caseChoose
Are the retrieved passages relevant to the employee's question?Context relevance
Do the retrieved passages contain all the information in the ground-truth text?Context coverage
Does the answer avoid stating things that are not supported by the retrieved passages?Faithfulness
Is the answer accurate?Correctness
Does the answer address every part of a multi-part question?Completeness
  • Relevant retrieval, grounded answers and factual correctness are separate measurements.

Guardrails policy types

Requirement / use caseChoose
Block any conversation about cryptocurrency investment recommendations, however the user phrases the request.Denied topics
Detect account numbers and email addresses in responses and replace them with placeholders.Sensitive information filters
Block responses that contain the exact names of specific competitor banks.Word filters
Detect user attempts to override the system instructions (jailbreak attempts).Content filters (prompt attack category)
In a retrieval augmented generation (RAG) question-answering feature, flag answers that add facts not found in the retrieved passages.Contextual grounding checks
Validate that answers about loan eligibility are logically consistent with the bank's written lending rules.Automated Reasoning checks
  • Contextual grounding compares with a source; Automated Reasoning checks compare with formal rules.

Explainability vs interpretability vs transparency

Requirement / use caseChoose
Reviewers can trace how the model works internally, for example by reading the coefficients of a linear model.Interpretability
People can understand, in plain terms, why the AI system behaves as it does, even when its underlying model is too complex to inspect directly.Explainability
Stakeholders get the information they need, such as disclosure of AI use and known limitations, to make informed choices about engaging with the AI system.Transparency
A feature importance score that is valid for the model across all inputs.Global explanation
A feature attribution score that applies to a single model output.Local explanation
  • Disclosing limitations differs from explaining one prediction or inspecting model internals.

Governance and security services

Requirement / use caseChoose
Download AWS's ISO and SOC reports to give to the company's auditors.AWS Artifact
Continuously evaluate whether SageMaker AI endpoint configurations use AWS KMS keys, and flag noncompliant resources.AWS Config
Record which IAM principal deleted a Bedrock knowledge base and when.AWS CloudTrail
Continuously scan Lambda functions that call the model for known software vulnerabilities.Amazon Inspector
Get recommendations about open S3 bucket permissions, idle resources, and service limits.AWS Trusted Advisor
  • CloudTrail answers who did what; Config checks resource configuration; Artifact provides AWS reports.

Anonymization vs pseudonymization vs differential privacy

Requirement / use caseChoose
Replace student identifiers with consistent placeholder values and keep the mapping in a separate, protected store so that authorized staff can re-identify recordsPseudonymization
Irreversibly remove or generalize identifying details so that the data no longer relates to an identifiable personAnonymization
Add a calibrated amount of statistical noise to query results so that no individual's data can be inferredDifferential privacy
Make stored training files unreadable to anyone who does not have access to the keyEncryption at rest with AWS KMS keys
  • Encryption protects access but does not by itself make personal data anonymous.

AI governance frameworks and documentation

Requirement / use caseChoose
An international, certifiable standard that specifies requirements for an Artificial Intelligence Management System (AIMS).ISO/IEC 42001
A voluntary U.S. framework to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI systems.NIST AI Risk Management Framework (AI RMF)
An AWS mental model that classifies generative AI use cases from least to greatest ownership of the model and data.Generative AI Security Scoping Matrix
A set of AWS best-practice questions that helps builder teams build and operate AI solutions responsibly.AWS Well-Architected Responsible AI Lens
AWS-published documentation of an AWS AI service's intended use cases, limitations, and responsible AI design choices.AWS AI Service Cards
  • A framework, a certifiable management-system standard, and an AWS service card are different artifacts.